CHILDREN AND FAMILIES
EDUCATION AND THE ARTS
ENERGY AND ENVIRONMENT
HEALTH AND HEALTH CARE
INFRASTRUCTURE AND
TRANSPORTATION
INTERNATIONAL AFFAIRS
The RAND Corporation is a nonprofit institution that
helps improve policy and decisionmaking through
research and analysis.
This electronic document was made available from
www.rand.org as a public service of the RAND
Corporation.
LAW AND BUSINESS
NATIONAL SECURITY
Skip all front matter: Jump to Page 16
POPULATION AND AGING
PUBLIC SAFETY
SCIENCE AND TECHNOLOGY
TERRORISM AND
HOMELAND SECURITY
Support RAND
Purchase this document
Browse Reports & Bookstore
Make a charitable contribution
For More Information
Visit RAND at www.rand.org
Explore theRAND National Defense
Research Institute
View document details
Limited Electronic Distribution Rights
This document and trademark(s) contained herein are protected by law as indicated
in a notice appearing later in this work. This electronic representation of RAND
intellectual property is provided for non-commercial use only. Unauthorized posting
of RAND electronic documents to a non-RAND website is prohibited. RAND
electronic documents are protected under copyright law. Permission is required
from RAND to reproduce, or reuse in another form, any of our research documents
for commercial use. For information on reprint and linking permissions, please see
RAND Permissions.
This report is part of the RAND Corporation research report series.
RAND reports present research findings and objective analysis that
address the challenges facing the public and private sectors. All RAND
reports undergo rigorous peer review to ensure high standards for research quality and objectivity.
N AT I ONAL DEFENSE R ESE AR C H IN S TITUTE
Using Behavioral
Indicators to Help Detect
Potential Violent Acts
A Review of the Science Base
Paul K. Davis, Walter L. Perry, Ryan Andrew Brown, Douglas Yeung,
Parisa Roshan, Phoenix Voorhies
Prepared for the United States Navy
Approved for public release; distribution unlimited
The research described in this report was prepared for the United States
Navy. The research was conducted within the RAND National Defense
Research Institute, a federally funded research and development center
sponsored by the Office of the Secretary of Defense, the Joint Staff,
the Unified Combatant Commands, the Navy, the Marine Corps, the
defense agencies, and the defense Intelligence Community under Contract
W74V8H-06-0002.
Library of Congress Cataloging-in-Publication Data
Davis, Paul K., 1952Using behavioral indicators to help detect potential violent acts : a review of the science
base / Paul K. Davis, Walter L. Perry, Ryan Andrew Brown, Douglas Yeung, Parisa
Roshan, Phoenix Voorhies.
pages cm
Includes bibliographical references.
ISBN 978-0-8330-8092-9 (pbk. : alk. paper)
Terrorism—Prevention. 2. Behavioral assessment. 3. Psychology—Methodology.
I. Title.
HV6431.D3268 2013
363.325'12--dc23
2013024014
The RAND Corporation is a nonprofit institution that helps improve policy
and decisionmaking through research and analysis. RAND’s publications do
not necessarily reflect the opinions of its research clients and sponsors.
Support RAND—make a tax-deductible charitable contribution at
www.rand.org/giving/contribute.html
R® is a registered trademark
Cover photo by Karl Baron via flickr
© Copyright 2013 RAND Corporation
This document and trademark(s) contained herein are protected by law. This
representation of RAND intellectual property is provided for noncommercial use only.
Unauthorized posting of RAND documents to a non-RAND website is prohibited.
RAND documents are protected under copyright law. Permission is given to duplicate
this document for personal use only, as long as it is unaltered and complete. Permission
is required from RAND to reproduce, or reuse in another form, any of our research
documents for commercial use. For information on reprint and linking permissions,
please see the RAND permissions page (www.rand.org/pubs/permissions.html).
RAND OFFICES
SANTA MONICA, CA • WASHINGTON, DC
PITTSBURGH, PA • NEW ORLEANS, LA • JACKSON, MS • BOSTON, MA
DOHA, QA • CAMBRIDGE, UK • BRUSSELS, BE
www.rand.org
Summary
Government organizations have put substantial effort into detecting
and thwarting terrorist and insurgent attacks by observing suspicious
behaviors of individuals at transportation checkpoints and elsewhere.
Related technologies and methodologies abound, but their volume and
diversity has sometimes been overwhelming. Also, effectiveness claims
sometimes lack a clear basis in science and technology. The RAND
Corporation was asked to review the literature to characterize the base
in behavioral sciences relevant to threat detection, in part to help set
priorities for special attention and investment.
Purpose and Approach
Our study focused on the science base for using new or nontradi­
tional technologies and methods to observe behaviors and how the
data gathered from doing so might—especially when used with other
information—help detect potential violent attacks, such as by suicide
bombers or, as a very different example, insurgents laying improvised
explosive devices (IEDs). Behavioral indicators may help identify individuals meriting additional observation in an operational context such
as depicted in Figure S.1. For that context, security personnel at a checkpoint are assessing (blue oval) whether an individual poses some risk
in the limited sense of meriting more extensive and perhaps aggressive
screening, follow-up monitoring, or intercept. They obtain information directly, query databases and future versions of information-fusion
centers (“pull”), and are automatically provided alerts and other data
xiii
xiv
Using Behavioral Indicators to Help Detect Potential Violent Acts
Figure S.1
A Contextual View of the Detection Effort
Proximate behavioral and
other information (e.g.,
near checkpoint or
possible target)
Assess risk of individual
or small group
Queries
Alerts
Data,
fusion centers
Concern? Decision on
secondary screening,
monitoring . . .
Previous
information
(watchlist,
individual-level
data, . . .)
Information for
next time
RAND RR215-S.1
(“push”). They report information that can be used subsequently. In
some cases, behaviors of a number of individuals over time might suggest a potential ongoing attack, even if the individuals are only pawns
performing such narrow tasks as obtaining information.
Figure S.1 refers to “other information” (top left). Although our
study is concerned with detecting imminent threats rather than gathering broad information for internal security or intelligence, such
information—perhaps accumulated over years—can play an important role. Where might that information be found, how might it be
structured, and what indicators might be involved? We focus on what
may be possible technically, without analyzing tradeoffs with privacy
and civil liberties. We do, however, note troublesome issues raised by
the technology and methods, point readers to an in-depth study of such
matters by the National Academy of Sciences, and suggest research on
ways to mitigate the problems.
Figure S.2 shows relationships among our key constructs. A base
of technology and methods (left) allows detecting behavioral indicators (bottom right). Moving upward, these give signals about activities,
which are grouped into activity classes called phases. Analysis can then
assess whether the totality of information (including nonbehavioral
Summary
xv
Figure S.2
Relationships Among Constructs
Other
information
Basis for concern
Phases
Technology and
methods for
observation and
analysis
Activities
Behavioral indicators
of activities
Overall assessment, from
indicators, of need for follow-up
on individual or small group
Aggregate activities providing
evidence (perhaps weak or
ambiguous) that raises concern
about subject
Activities providing evidence
(perhaps weak or ambiguous)
that raises concern about
subject
Observables providing evidence
of potentially threatening
activity
RAND RR215-S.2
information) adds up to a basis for concern justifying more screening,
monitoring, precautionary defensive measures, or preemptive action.
Since detecting a “basis for concern” (i.e., need for further checking)
will probably have a high false alarm rate, a system using this approach
must be efficient and reasonable if it is to be acceptable.
Figure S.3 is a conceptual model showing phases within which
lower-level activities occur. The model merely identifies where to look
for information. As indicated at the bottom of the figure, there are
many possible indicators of the activities and a number of technologies
and methods to use in observing them. The model is merely heuristic rather than a rigorous decomposition or timeline. Activities may
be performed by multiple individuals, not occur, or occur in different
order. Some activities could occur in more than one phase.
Figure S.4 uses the “Developing Intent” phase of Figure S.3 to
illustrate how phases, activities, and indicators relate to technologies
and methods. For each of three activities, Figure S.4 shows potential indicators. The lowest box shows some relevant technologies and
methods. The Developing Intent phase is unusual in that it includes
xvi
Figure S.3
Conceptual Model of Opportunities for Observation
Nonbehavioral data
Phases (need not all be sequential)
Activities
Developing
Intent
Psychological
development
Target
identification,
ISR
Recruitment
or joining
Materiel
acquisition,
testing and
development
Aftermath
Execution
ISR
Development
of strategic
priorities
Psychological
convergence
Immediate
Pre-Execution
Planning and Laying
Groundwork
Deployment,
positioning
Psychological,
physiological
preps
Other
long-lead
preparations
Training,
mission
rehearsal
CONOPs
development
Post-attack
Adaptation
reporting
Strategic
Protective
communication measures
Logistical
preparation
Changed
social
patterns
Ritual practices
Deception,
concealment
Coordination,
communication
Main
attacks
Target
shaping,
feints
Indicators (nominally of specific activities, but potentially relevant to others as well, or to an overall judgment)
Technologies and methods
NOTES: ISR = intelligence, surveillance, and reconnaissance. CONOPs = concept of operations. Indicator-phase connections are
nominal. Some activities can occur in more than the phase indicated. Associations of activities with phases are nominal. Some
activities can occur in additional phases.
RAND RR215-S.3
Using Behavioral Indicators to Help Detect Potential Violent Acts
Evidence of violent intent
Summary
xvii
Figure S.4
Illustration of Methodology
Activities
• Act so as to show
hatred, prejudice,
trauma, or shame
• Exhibit strong
motivation for
prestige glory, status
• Demonstrate
intolerant black-andwhite thinking
• Indicate perception
of ”rightful place”
or status, but with
meager resources
and attainment
Technology
and methods
Psychological
and emotional
development
Behavioral indicators
Develop intent
Psychological
convergence
Recruitment
or joining
• Approach life and
decisions through
ideological framework
• Proselytize; reinforce
practices and beliefs
in others
• Indicate reinforced
commitment
• Seek out materials
with radical content
• Change behavior at
school or home
• Explore different
organizations and
strategies
• Attend radical
mosques or events
• Form new connections
in social network to
known terrorist
elements
• Adhere strictly to
organizational
standards
• All-source, real-time data fusion
• Whole-life data
• Online content “scrapers”
• Social network analysis
• Mobile device tracking
• Automated textual analysis of communications or online content
RAND RR215-S.4
early-in-life activities, such as might be observed by parents, neighbors,
teachers, physicians, local law enforcement, and others long before an
individual becomes involved in anything violent. In the main text, we
discuss the phases separately, identifying generic activities and potential indicators.
xviii
Using Behavioral Indicators to Help Detect Potential Violent Acts
Technology and Methods
We found it useful to highlight technologies and methods in three
cross-cutting classes of data: (1) communication patterns, (2) “patternof-life” data, and (3) data relating to body movement and physiological
state. Most of the methods suffer from signal-to-noise problems and
false alarms; some are vulnerable to countermeasures.
Communication Patterns
Communications occurs in, e.g., face-to-face meetings, Internet chat
rooms, and cell phones. Large commercial and intelligence-sector
investments have yielded techniques to monitor and analyze these
communications, which we also treat in three groups: online communications and analysis, text analysis and natural-language processing,
and speech analysis.
Online Communications. Using data collected by online-content
“scrapers” and subsequent human efforts, it is sometimes possible to
infer motivations, emotions, and ideologies from online statements
and actions (e.g., as with the New York Police Department following Twitter posts such as “people are gonna die like Aurora” (referring to the July 20, 2012, movie-theater shooting in Aurora, Colorado). Real-time social-media search tools can help monitor and track
discussions of potential targets. Keystroke loggers and other programs
can reveal past and current searches for materiel; registrations or payments to training programs; location information; and information on
past searches. Changes of activity may occur when terrorists “go dark”
before an attack, when logistical preparations for an attack are intense,
or when calls for vengeance arise after an event such as Osama bin
Laden’s killing.
This and all the methods include false alarms (e.g., users may have
benign reasons for their purchases or mere curiosity as they investigate
troublesome websites), low signal-to-noise ratio, and vulnerability to
such counters as burying information amidst innocuous communication or using anonymous or false accounts.
Text Analysis and Natural­Language Processing. Techniques to classify texts and analyze content are fairly well developed, although less so
Summary
xix
with respect to emotion and intent. Explicit content may include bragging, ideological statements, or admiration for terrorist leaders. Textual
analysis of style can detect word-usage patterns associated with typical attacker motivations and such emotions as anger, humiliation, and
shame. Style of communication does not depend on content, i.e., on
specific topic, and can suggest relationships and status within a social
network. Textual analysis of style has been used for, e.g., detecting corporate fraud, terrorist interrogations, and criminal testimony. Naturallanguage processing can analyze massive amounts of text about which
little is known in advance, classifying documents to be analyzed further
by subject-matter experts. Clustering methods can identify concepts
and such topics as weapons, tactics, or targets. Such mathematical techniques as latent semantic indexing can help understand concepts and
have the advantage of being language-independent. Machine translation can often turn foreign-language texts into something analyzable
without foreign-language expertise or language-specialized software.
Speech-recognition technology can greatly increase the amount of text
available for text analysis. It can also help identify individuals.
A primary shortcoming is nonspecificity; that is, detected patterns (even if apparently threatening) may be unrelated to any imminent threat, and their interpretation often depends on cultural and
individual idiosyncrasies. A shortcoming of the research base itself
is that much linguistic-style analysis has been done only on archival
data; more testing and validation is needed with “real-life” data sets.
Top researchers caution against expecting highly reliable detections or
interpretations and suggest the need for very large data sets that reveal
many cultural and individual differences.
Speech Analysis of Content. Several robust indicators exist for connecting vocal content and narratives with lying and deception. These
include the subject (1) distancing himself from untruthful statements
by, e.g., using the third person or otherwise seeming less verbally
involved; (2) issuing discrepant statements; (3) providing less detail;
(4) exhibiting less logical structure and less subjectively plausible stories; (5) providing less context; and (6) making less spontaneous corrections or claiming lapses of memory.
xx
Using Behavioral Indicators to Help Detect Potential Violent Acts
This approach’s primary shortcoming in assessing deceptive or
hostile intent is that interpreting lexical and vocal indicators of lying
and deception depends on context, individual variability, and appreciation of nonthreatening explanations. Optimally, analysis has data
on the individual’s speech in a normal nondeceptive/nonhostile state.
Where this is infeasible, the potential increases markedly for failed
detections and intolerably many false alarms. Table S.1 summarizes
Table S.1
Considerations and Caveats: Detection and Analysis of Communication
Patterns
Domain
Status
Shortcomings
Measurement
and
Upside Potential Requirements Vulnerabilities
Online
Extensive
Given trends,
communication collection
even more
and activities
and analysis
and varied
occurs today
interactions
for commercial
will be
and intelligence
available for
reasons.
collection.
Technologies and
methods for
analyzing such
online activities
are still relatively
unproven in
either academic
or operational
settings.
Text analysis
and naturallanguage
processing
A considerable
research base
exists with
numerous past
applications.
Even naturallanguage
processing can be
highly accurate
in specific
experimental
settings.
Tools already
Methods have
exist. However, not been well
challenges
validated in
for dealing
academic or
with massive
operational
volumes of
settings.
noisy data are Low signal-toformidable.
noise ratio.
Effects of
encryption,
using “code,”
using
anonymizers,
or moving
offline.
Using
Online text
operational
is naturally
data to train
occurring
and to create
and publicly
baselines
accessible,
could improve requiring
detection of
only passive
deception,
collection.
hostility, or
Active
extremist
elicitation of
patterns.
text or oral
Naturalstatements
language
is possible in
techniques,
some security
given training
contexts,
sets, could
such as
quickly analyze checkpoints or
large amounts interrogations.
of data.
Context and
cultural
dependence.
Inadequate
testing in
operational
settings.
Need for
substantial
data.
Summary
xxi
Table S.1—Continued
Domain
Status
Speech
This has been
analysis: lexical validated in
and vocal cues laboratory
settings,
including those
specific to
counterterrorism.
Shortcomings
Measurement
and
Upside Potential Requirements Vulnerabilities
Advances in
Such analysis
Physiological
protocols
currently
drivers, such
for rapid
requires
as anxiety
assessment
skilled security and changes
of speech
personnel
in vocal tone,
patterns and
asking
are individualcontent would questions
dependent.
have wide
and making
May be subject
applicability
judgments.
to counters,
for screening,
especially if
checkpoint, or
criteria for
other situations
judging are
involving
known.
conversations
with security
personnel.
results of our review for the assessment of communication patterns and
content.
Pattern-of-Life Data
It is possible to analyze patterns of communication, travel, purchasing,
and other matters using existing records and databases (many held by
private industry). We discuss mobile-device tracking, using existing
records, and machine learning for pattern detection. These raise profound social questions about what kind of data can and should be collected and analyzed.
Mobile­Device Tracking. Ubiquitous mobile devices provide a
wealth of data on personal information, social relationships and networks, locations, patterns of movement and interactions, preferences,
political opinions, the spread of information, and patterns of how opinions and preferences change. Also, mobile-device usage is related to the
“Big Five” personality traits (openness, conscientiousness, extraversion,
agreeableness, and neuroticism).
One shortcoming of such data is that much social networking
through mobile devices is increasingly “muddy” and in many cases
divorced from both intent to meet and intent to act in the offline “real
xxii
Using Behavioral Indicators to Help Detect Potential Violent Acts
world.” This complicates inference-making about patterns of communication and their connection to actual threat. And, of course, people
can go offline.
Existing Records. It is sometimes possible to develop individual
profiles from information about, e.g., experiences, behaviors, and relationships over time, and to provide context for assessing other incoming data. The data could come from school records, criminal records,
interrogation reports, and so forth. Additionally, surveillance cameras
are now common in public and business settings, allowing for the possibility of tracking individual patterns-of-life. Integrating such data
requires analytic techniques, including those for all-source, real-time
big-data fusion. Related analytic tools are increasingly available from
such providers of cloud computing as Google and Amazon and socialmedia companies.
The shortcomings include, of course, the administrative, jurisdictional, legal, and database challenges of extracting and combining
data across multiple sources and owners within and outside the United
States. Accuracy matters for this type of analysis, whereas commercial
applications often do not require high accuracy to improve the targeting of marketing efforts.
Machine Learning and Big­Data Analysis. Given the sheer magnitude of data, it is increasingly important to analyze information without the benefit of prior hypotheses or known points of comparison.
“Supervised” machine-learning techniques use known data sets to
train the algorithms, which can then classify data, identify relationships, and discover concepts. “Unsupervised” learning proceeds without the aid of such known prior knowledge. It seeks to find structure
in the unlabeled data set. For example, researchers have used thousands
of YouTube images for unsupervised detection of high-level features
such as faces. Potentially, such techniques could recognize images suggesting imminent threat. Such machine-learning techniques have been
applied to uncover fraud, to recognize deception in computer-mediated
communication, and for predictive policing. Artificial neural-network
models are promising and can be applied in real time. Video or image
analysis and machine-learning techniques could be employed to find,
Summary
xxiii
for example, such activities as shaving heads and prayer activities in
martyrdom videos.
One shortcoming is that machine-learning techniques often
require a large amount of data. At least in the public domain, sufficiently large databases of violent attacks and other events do not exist
for topics such as terrorism. One innovative method for obtaining large,
labeled data sets is to “crowd-source” collecting and labeling individual
pieces of information.
Table S.2, which is analogous with Table S.1, is our assessment of the various approaches focused on records-based whole-life
information.
Indicators from Physical Movement and Physiology
Behavioral science has identified many nonverbal behaviors that are
statistically associated with emotional and psychological state and
with deception or violent intent. These can be roughly categorized into
(1) kinetics (including gross motor movements) and (2) observation of
physiological state. As discussed in the main text, many findings are
controversial among scientists, and between scientists and operators,
but our summary assessment follows.
Kinetics and Gross Movement. Existing technology can collect
data on kinematic patterns (movement). Surveillance and reconnaissance platforms (e.g., tower cameras or drone systems) can monitor
individuals as they maneuver before an attack. Video systems can view
individuals before attacks and collect information on individuals who
frequent potential attack sites, providing a baseline for identifying individuals engaged in pre-execution activities. For example, the gait of
people who may be carrying weighted objects, such as IEDs, may be
compared against baseline “gait signatures.” Existing recordings of terrorism incidents can provide data for setting parameters on new analysis tools. The Defense Advanced Research Projects Agency (DARPA)
has funded biometric technologies for identification at a distance and
for early warning. Another approach seeks to automate recognition of
potentially threatening body postures or poses.
Incorporating emotion into machine-learning methods may
increase their future utility. To do so, “affective computing” may need
xxiv
Using Behavioral Indicators to Help Detect Potential Violent Acts
Table S.2
Considerations and Caveats for Pattern-of-Life Data
Domain
Status
Mobile- Algorithms
device
to predict
tracking individual
movement
patterns,
preferences,
etc., have been
developed and
validated in
laboratory and
experimental
settings, but can
benefit from
more naturalistic
validation.
Upside Potential
Mobile devices
will continue to
add connectivity
features
that enable
tracking (e.g.,
location and
motion sensors,
Near Field
Communication
chips).
Measurement
Requirements
Shortcomings and
Vulnerabilities
Mobile device
Not traveling with
tracking
or turning off
may require
device will defeat
device-owner
methods based
permissions or
on mobile-device
cooperation of
whereabouts.
communications Mobile-to-mobile
network
communication
providers.
is often divorced
from “real-life”
behaviors and
intent.
Pattern- Validating
Pattern-of-life
Measurement
Pattern-of-life
of-life
techniques to
data may allow
does not
data may be
data
analyze large
integrating
require active
vulnerable to
amounts of
disparate data
or voluntary
“cover” activities
pattern-oftypes to build
consent.
and behaviors.
life data may
fuller behavioral However, access Databases and
be difficult
profiles on
to various
algorithms
in academic
individuals
databases held
for detecting
settings.
of interest.
by commercial
threatening
Commercial
Accessing and
or private
patterns
data sets and
integrating data sources may be
are in early
analytic tools
is an issue.
necessary.
development.
are increasingly
available.
Machine Machine-learning Machine-learning
learning techniques
and big-data
and
have been
analysis may
big-data extensively used “discover”
analysis
and validated
unknown
in experimental patterns or
and some
activities hidden
applied settings. in large amounts
Such techniques of data, but
have been used massive amounts
in national
of data are
security and law needed for
enforcement.
training.
Measurement
Learning
does not
techniques are
require active
probabilistic and
or voluntary
vulnerable to
consent.
noisy data.
A large amount Current systems
of data or
do not
a strong
understand how
hypothesis
to associate
regarding
behaviors
relevant activity of multiple
is required.
threatening
individuals.
Summary
xxv
to select from various psychology and neuroscience findings and theories of emotion (e.g., “appraisal models”). Often, subsystems of monitoring and interpretation of stimuli can be computationally modeled.
Improvements are possible when distinguishing between emotional
states that differ in arousal, such as anger and sadness. Methods being
developed to analyze gait signatures could be applied to such existing
commercial technologies as cameras used for the Microsoft Kinect and
Nintendo Wii game systems’ motion-capture capability.
Human observers and analysts may also be employed, but results
depend on such factors as training, individual talent, and observer
bias. Detecting deceptive movements is easier for people experienced in
employing the same deceptive movement patterns. People are best able
to detect emotions associated with gait when the human walkers are
expressing anger. Inference from merely a single stride can be highly
accurate, suggesting that gait can be used to recognize affect. Performance varies by individual, and women may be better than men at
determining actions and recognizing emotions from movements such
as walking.
Analysis of kinetics and gross motor movements should apply to
a wide variety of security contexts, although validation in naturalistic
settings is needed and, as often occurs in looking for behavioral indicators, the indicators may arise for benign reasons, such as people being
anxious at security screenings or checkpoints.
One challenge for gait analysis is that current detection systems
and protocols are often built using simulated behaviors (e.g., with
actors). More naturalistic (real-world) observations are needed.
Physiological State and Reactions. Observing physiological state
holds promise for detecting deception and other behaviors. We touch
upon polygraph testing, other measures of peripheral nervous system
response, electroencephalograms (EEGs), vocal stress, and facialexpression analysis.
Polygraph testing has long been employed and found useful as
part of an investigatory process (particularly because people often
“open up” in the process), but is not by itself reliable. A great divide
exists between the bulk of the academic community, who remain quite
skeptical, and the community of “operators,” who insist on its useful-
xxvi
Using Behavioral Indicators to Help Detect Potential Violent Acts
ness as one tool in a process. Newer approaches using some of the same
physiological signals as in polygraphs (heart rate, blood pressure, etc.)
are in development with respect to detection of potential deception or
hostile intent.
New technologies using electroencephalograms (EEGs) allow
some physiological features to be observed without “wiring up” individuals, sometimes at a distance, and sometimes covertly or surreptitiously, as with using heat-sensitive cameras to detect capillary dilation
and blood flow to the face and head. There is some evidence of unique
value in indicating deception or imminent action by an individual if
baseline information is available for that specific individual ahead of
time or if credible intelligence about a possible attack is available. Most
of the technologies are in a relatively early stage of development, but
there does seem to be potential. Measurement of physiological signals
closer to the central nervous system (i.e., the brain) holds the most
promise for detecting guilt and behavioral intent.
Evidence of vocal tension and higher vocal frequency may also
be predictors of stress and deception, and a few observable aspects of
speech are much more difficult for an individual to control than other
indicators of deception, but countermeasures that obscure differences
from the baseline of normalcy are definitely feasible.
Humans appear to share universal facial expressions indicative
of underlying emotional and motivational states. Cultural differences
seem to affect only secondary aspects of facial expressions. The seven
fundamental emotions—anger, disgust, fear, happiness, sadness, surprise, and contempt—are displayed on the face with some fundamental features that are generally recognizable on all humans (barring
neurological impairment). For the purposes of detecting pre-incident
indicators, the most promising domain of facial expression analysis
involves facial micro-expressions—involuntary expressions of emotion appearing for milliseconds despite best efforts to dampen or hide
them. Whether the relevant behavior is smuggling weapons, traveling
on forged documents, or hiding anger or anxiety near security officials,
facial micro-expressions can be important indicators of deception or
some kind of mal-intent.
Summary
xxvii
At least currently, the two primary problems with using physiological indicators are (1) nonspecificity (the indicators may stem from
many causes, most of them benign) and (2) individual differences (the
observables that indicate attack or deception differ markedly across
individuals, which requires establishing sound individual-centered
baselines). Countermeasures are a problem with polygraphs, but perhaps less so with EEG methods. Even with polygraphs, empirical results
have varied. Some drugs, for example, have not reduced detection rates
as expected, but physical training can be effective as a countermeasure. Controlling vocal stress indicators is difficult, but countermeasures can obscure distinctions between baseline and stressed behavior.
Facial expressions suffer from the same problems of nonspecificity, but
they have the advantage of being more closely linked to motivational
state and intent than are other physiological signals. Individual differences are also important: A psychopathic attacker, for example, might
be more likely to show micro-expressions of “duper’s delight” while
passing through a checkpoint undetected, while a nonpsychopathic
attacker might instead show micro-expressions of fear (as would a perfectly harmless nervous traveler).
While the link between micro-expressions and deception is well
evidenced, utility in security-related settings is another matter. Coding
emotional expressions currently involves hours of labor to analyze seconds of data, making this technique unsuitable for use in real time
at checkpoints or other screening areas. However, a training system
appears to increase the capacity of individuals to detect facial expressions and micro-expressions with demonstrated evidence of effectiveness in clinical populations.
Recognition of emotional expressions based on automated algorithms and computation is still in its infancy, but this is an active field
of development, and improved algorithms are likely to yield greater
accuracy and robustness. Furthermore, as with many pre-incident
indicators of attack, emotion-recognition algorithms that fuse multiple parameters seem to perform much better than inferring emotional
state simply from facial expressions alone.
Of course, checkpoints or other security environments are
dynamic locations where it is difficult to capture high-resolution video
xxviii
Using Behavioral Indicators to Help Detect Potential Violent Acts
(or audio or physiological data) for individuals, but such detailed information is often necessary for effectiveness.
Table S.3 is our assessment of how the approaches based on
detecting intent from physiological indicators stand in terms of maturity, potential, measurability, and vulnerability to countermeasures.
Cross-Cutting Themes
A number of cross-cutting issues arose in our review. These suggest a
notional framework for thinking about detection systems. Although
the relevant metrics have by no means been defined as yet, much less
metrics that take into account cost-effectiveness, a goal for future
analysis might be to place something like the framework shown in
Figure S.5 on a solid scientific and analytic basis. Although it is surely
not yet “right” or well defined, this framework conveys a sense of what
is needed for sounder discussion. Further, despite its shortcomings, we
Figure S.5
A Notional Framework for Characterizing an Overall System
Approximate layering
Mitigation
of false-alarm
consequences
100
80
60
Subsystems’ sensitivity,
selectivity (no stimulation)
Primitive
Baseline
Advanced
Very advanced
40
20
0
Information
fusion
Observation distance,
covertness, automaticity
RAND RR215-S.5
Sensitivity and
selectivity with
behavioral stimulation
Countermeasure
resistance of subsystems
Table S.3
Detecting Hostility or Deception from Movement Physiology and Movement
Domain
Status
Upside Potential
Measurement Requirements
Shortcomings and
Vulnerabilities
Physiological Indicators have been
state and
validated in laboratory
reactions
and experimental
settings, with some
experimental paradigms
simulating elements of
counterterrorism and
some (facial) cutting
across culture.
In some cases (e.g.,
voice stress and facial
indicators), automated
recognition shows
potential but currently has
high error rates.
Currently, measurement
Differences across contexts
requires direct application
and individuals.
of sensors or the physical
Nonspecificity.
observation of, e.g., facial
Influence of drugs and
flushing, sweating.
training (e.g., to dampen
Some (e.g., facial) require
or obscure differences
lighting and proximity with
between baseline and
currently painstaking coding
signals).
feasible only for high-value
Masking,
in some cases
interrogations. Success requires
exceptional “natural” talent or (e.g., sunglasses or plastic
surgery)
training, but limited available
data suggests training is
Some differences exist
effective.
(perhaps not critical) across
culture.
Measurements are most
valuable when comparing
Masking (e.g., sunglasses,
against an individual’s
plastic surgery, or Botox
baseline, which is only feasible for facial), but this may
in voluntary monitoring or
also be an indicator.
interrogation context.
Internal physiological
reactions are relatively
automatic and difficult
to control (e.g., micro
tremors in speech or micro
facial expressions).
Probing of various sorts
(even seemingly random
conversations) can trigger
reactions.
Certain elements of
facial expression are
very difficult to alter
voluntarily, including
micro-expressions.
xxix
Some security contexts may not Masking with deceptive
allow for sufficient physical
movements.
movement to be interpretable Sensitivity to context and
(e.g., interrogation).
individual differences.
Nonspecificity: triggering
by diverse emotions and
motivations.
Summary
Kinetics and Indicators have been
Gross motor movements
gross motor validated for human
may reveal action,
movement
observation and
intent, or deception.
automated analysis
On-foot motions may
in laboratory and
be unavoidable in such
experimental settings,
proximal security settings
including some
as checkpoints.
operational settings (e.g., Gross motor movement
for gait of individuals
may be passively
carrying weighted
observed, but also actively
objects).
elicited.
xxx
Using Behavioral Indicators to Help Detect Potential Violent Acts
have found the framework qualitatively useful for discussing the issues
that arose in our critical survey.
Figure S.5 uses a radar/spider plot to characterize a given detection system along seven dimensions, with a score of 100 corresponding to a system that has been optimized along that dimension while
considering feasibility and cost-effectiveness. The score given to a lesser
system is a rough and subjective characterization of how much has
been accomplished relative to what would be accomplished optimally.
The dimensions relate to (1) appropriate layering, (2) sensitivity and
selectivity of subsystems for information when it is obtained and countermeasures are absent, (3) behavioral stimulation, (4) countermeasure
resistance of those subsystems, (5) the ability to obtain the information in desirable ways that may include automated observations from
a distance, perhaps without subjects being aware of the observations,
(6) information fusion, and (7) mitigating the consequences of false
alarms when they occur.
These dimensions relate to overall system effectiveness as shown
in Figure S.6, which highlights the need for both effective detection
(minimizing “false negatives”) and management of the false-alarm
problem (minimizing “false positives” or mitigating their negative consequences). Returning to Figure S.5, we see that it illustrates notionally
Figure S.6
Factors Affecting Overall System Effectiveness
Detection Effectiveness
6. Information
fusion
1. Appropriate
layering
3. Behavioral
stimulation
RAND RR215-S.6
Effectiveness in Limiting
False Alarms and Their
Consequences
4. Countermeasure
resistance
7. Mitigation of
false-alarm
consequences
5. Observation 2. Subsystem sensitivity and selectivity
distance,
(no stimulation or countermeasures)
automaticity,
etc.
Summary
xxxi
what progress might look like over time. For the example, we characterize the baseline (today’s system) as more advanced in some respects
than others, with layering having been taken seriously for some time,
but with information fusion, for example, being relatively primitive
and with too little work having been done to mitigate the consequences
of false alarms when they occur. Progress would correspond to systems with contours farther and farther toward the extremity of the
radar/spider plot (cautions in interpreting such plots are discussed in
the main text). Subsequent paragraphs touch on each dimension.
Appropriate Layering
The value of layering in detection systems is well discussed elsewhere
and merits little discussion other than to note that the fatal flaw in
some assessments is assuming that the various layers are independent,
which is not the case when, for example, they share lax or incompetent management. How much layering is appropriate depends on many
factors.
Subsystem Sensitivity and Selectivity
Screening can be based on many types of information, such as background checks, overtly observable characteristics (including the carrying of weapons), or behavioral cues. Ongoing research is a mix of
laboratory- and field-based empirical research and modeling. As an
example, the Department of Homeland Security’s (DHS’s) Screening of
Passengers by Observation Techniques (SPOT) Program was designed
to help behavioral detection officers (BDOs) identify persons who may
pose a potential security risk at Transportation Security Administration (TSA)–regulated airports. It focuses on behaviors and appearances
that deviate from an established baseline and that may be indicative
(imperfectly, to be sure) of stress, fear, or deception. BDOs may refer
some of those whom they observe to additional screening. DHS conducted a large randomized trial of SPOT effectiveness and reported
that the program resulted in the detection of illegal behaviors at a significantly higher rate than random selections, with a small false-alarm
rate. Another empirical effort, Project Hostile Intent (later renamed
Future Attribute Screening Technology [FAST]), included consider-
xxxii
Using Behavioral Indicators to Help Detect Potential Violent Acts
ation of hard-to-suppress micro-expressions as it sought remote, noninvasive, automated sensor technology capable of real-time detection
of deception.
Although not behaviorally oriented, a second class of screening
method should be mentioned. It is illustrated by the “Trusted Traveler”
concept, which screens for those who can be excluded from some secondary screening. A 2011 analytical review demonstrated considerable
promise, especially if measures are taken to deter attacker efforts to get
into the program; aspects of the concept are now operational.
A constant issue in screening is how to trade off detection rate
against false-alarm rate. Doing so should depend on context. During a
period of high alert, security personnel can use less discriminate behavioral (and other) cues if they have additional temporary resources.
During such a period, the public also tends to be more forgiving of
inconvenience and somewhat higher false-alarm rates are tolerable.
As mentioned earlier, measuring physiological responses is most
effective when measuring actual brain activity rather than downstream
effects such as flushing. For example, electroencephalogram (EEG)
measurements have shown high effectiveness in laboratory experiments to detect deception by subjects in mock terrorist attacks. Such
methods, however, require the cooperation or coercion of individuals
and expensive monitoring equipment as well as credible prior intelligence about the details of a potential attack. The approach appears
to be countermeasure-resistant, but particularly aggressive individuals show less of the response being monitored, which would reduce
real-world effectiveness. Nonetheless, the successes are a remarkable
advance. Numerous empirical studies have also been performed on
“downstream” responses, including use of polygraph methods and
remote detection of peripheral physiological signals. These and other
methods have shortcomings, such as nonspecificity and sensitivity to
the “base rate” (the fraction of observed individuals who present an
actual threat). If the base rate is very low, then false alarms are high.
Behavioral Stimulation
Probing to stimulate behavioral responses can sometimes improve
detection effectiveness significantly. The basic concept has long been
Summary
xxxiii
familiar to law enforcement and many tangible examples exist, partly
as the result of U.S. security officials learning from extensive Israeli
practice. More generally, probing refers to the intentional stimulating of behavioral responses, such as by verbal questioning, anxietyraising changes of procedure or process, subliminal stimuli, or tests
with polygraph or EEG equipment. Probing may be polite, unobtrusive, or “in-your-face.” Some probing can definitely improve detectionsystem results, but related experimentation and formal assessment has
not been pursued as far as it might be. Verbal provocation and human
assessment of verbal and behavioral responses can be effective in some
circumstances without the use of sophisticated or expensive biological
monitoring equipment. Israeli airport and other transit officials have
used such techniques for many years, apparently with success (some of
it deterrence-related). Subjective assessment of the plausibility of reasons given for traveling, or being at a certain location, along with the
consistency of stories over time together provide the best clues about
hostile or deceptive intent. Further research should address contextually distinct tradeoffs between benefits for detection effectiveness and
negative consequences for civil liberties, commerce, and the perceived
legitimacy of the security system.
Allowing for and Dealing with Countermeasures
Much of the literature and even more of the advocacy-related discussion focuses on detecting behavioral responses in the absence of
countermeasures, but countermeasures are in fact a big problem, and
vulnerability to countermeasures should be a prime consideration in
evaluating investment programs. That said, countermeasures often
are not employed, are attempted poorly, or themselves create indicators. Thus, a balance must be struck in analysis: Worst-casing could
eliminate valuable measures, but optimistic assumptions could waste
resources and divert attention from more promising methods. Unfortunately, net judgments are often made informally and ad hoc. Analysis could improve this situation.
xxxiv
Using Behavioral Indicators to Help Detect Potential Violent Acts
Observation Circumstances: Remoteness, Covertness, Automaticity
Many of the potentially attractive technologies and methods currently
depend on such relatively benign circumstances as close-up observation
by humans, sometimes with a need to minimize “noise” relevant to
detection systems. Operational value, however, will be much enhanced
by improved capabilities to make observations from a distance, automatically, and in some instances without the subjects being aware of
the observation. Progress is being made by active technology efforts on
all of these. Some of the efforts are benefiting from commercial and
law-enforcement-system investments in, e.g., ubiquitous security-video
recordings; supervised and unsupervised computer search of data,
including “big data”; and new analysis techniques, such as those used
in data mining.
The Potentially Critical Role of Information Fusion
We found nothing on the horizon that presented a “magic bullet” for
threat detection, raising the potential importance of effective information fusion. We reviewed quite a number of methods for combining
information, ranging from very simple to more sophisticated methods. Notably, some classes of fusion have long been used. Indeed, polygraph testing combines information from several types of physiological signal. However, we have in mind information fusion that also
combines information across activities and phases. We considered a
number of possibilities.
Heuristic and Simple­Model Methods include checklists and risk
indexes, which are especially suitable for on-the-scene security personnel. Checklists are common already and can be of two kinds, which
are sometimes referred to as positive and negative (but with different
authors reversing which is which). As examples, any indicator, if met,
might trigger additional screening; alternatively, if all indicators are
met, secondary screening might be minimized. Index methods (scoring methods) typically characterize a risk level by summing indicator
scores, or by computing a risk as the product of a likelihood and a
consequence, with a score exceeding a threshold triggering additional
screening. Significantly, good scoring methods often need to be nonlinear and should be empirically validated rather than ad hoc. We
Summary
xxxv
also consider more complex “simple methods,” such as scorecards and
conditional-indicator sets.
More sophisticated integration methods are likely necessary in
future information-fusion centers, which would try to incorporate
behavioral indicators to overcome serious signal-to-noise and falsealarm problems. Accordingly, we reviewed mathematical informationfusion methods that might be adapted and extended (these methods
are discussed in more detail in Appendix D). Bayesian updating is well
understood and widely applied in other domains, but its usefulness in
our context is limited by its demands for many subjective estimates of
conditional probabilities for which there are and will continue to be
an inadequate base, and by limitations of expressiveness. Some newer
methods are based on Dempster-Shaefer belief-functions, which distinguish between having evidence for a proposition (such as the malign
intent of someone observed) and having contrary evidence (i.e., of
innocence). Evidence for both can be high, whereas if the language
used were that of simple probabilities, a high probability of malign
intent would imply a low probability of innocence. Dempster-Shaefer
theory requires fewer subjective inputs. Ultimately, however, there are
several major shortcomings in using that approach as well.
A much newer approach, called Dezert-Smarandache (DSmT)
theory, has not yet been widely discussed and applied, but something
along its lines has promise because it deals specifically with combining
evidence from sources and sensors that produce imprecise, fuzzy, paradoxical and highly conflicting reports—precisely the type of reports
encountered. For example, it allows characterizing the evidence that
both A and B are true; that one or the other of A or B is true (but not
both); or the evidence that A is true and the evidence that A is not true.
We also reviewed, briefly, the relevance of “possibility theory,” various multi-attribute theories, “mutual information” (which builds on
the concept of information entropy), and Kalman filtering. The best
method(s) for this problem area are not yet certain, but our review may
help to generate fruitful research in this critical area.
xxxvi
Using Behavioral Indicators to Help Detect Potential Violent Acts
Mitigating Costs of False Alarms
As mentioned repeatedly, a major challenge in detection systems is the
tradeoff between false negatives (failure to detect) and false positives
(false alarms), known as Type I and Type II errors. An understudied
problem amenable to research is how the broadly construed cost of the
latter can be reduced—not just by reducing the false-alarm rate, but
also by mitigating such bad consequences of false alarms as wasting
people’s time, raising their fears, insulting their dignity, or invading
their privacy. We identify three classes of initiative: (1) improve system
effectiveness (a “no-brainer”); (2) reduce effects on dignity and perceived violations of civil liberties (e.g., by transparency, explanation,
fairness, apology, and compensation); and (3) deter abuse by those
within the security system. Progress on the latter two is highly desirable for broad societal reasons and has many precedents in law enforcement. The negative consequences of false alarms alienate people, who
are then less likely to cooperate, volunteer suspicions, and support the
security system.
A Core Issue in the Use of Behavioral Indicators
Many of the subjects reviewed in our study are extremely contentious. Some of the controversy is scientific, relating to whether various detection methods are scientifically sound (or, as some would have
it, pseudo-science). The issue is not straightforward, because detecting
attacks by subjects such as terrorists involves looking for weak signals
amidst a great deal of noise in circumstances in which the “base rate” is
extremely low. The consequences of detection failure are very high, but
there are also profound negative consequences related to false alarms,
as mentioned above.
We could not resolve the controversies in this study, but Table S.4
makes distinctions useful in discussion. It compares how various methods that use behavioral indicators can be used. All of them have deterrent or cost-imposition value (second column). Would-be attackers
often fear the technology and methods and behave accordingly. All of
the methods can, when properly used and in proper circumstance, be
Summary
xxxvii
Table S.4
Some Comparisons of Where Behavioral Methods Have Value
Method
Flagging for Further Flagging with
Routine Screening
Prejudice for
Extended
Deterrence
Checking
Basis for
or Cost
and
Tool in
Arrest or
Imposition Automatic Human
Detention Interrogation Conviction
Polygraph
Yes
No
Yes
Maybe
Yes, but
No
Voice
stress
analysis
Yes
Yes
Yes
No
Yes, but
No
Facial
expression
Yes
Technology
not well
developed
Yes
No
Yes, but
No
EEG
Yes
Technology
not
developed
Yes
Maybe
Yes, but
No
Text or
speech
content
Yes
Maybe
Yes
Maybe
Yes, but
Maybe
Gait
analysis
Yes
Yes
Yes
Maybe
No
No
useful in providing incremental evidence on which subjects merit closer
scrutiny (third and fourth columns), although there are big variations
in whether they can be used automatically, remotely, and covertly. All
of the methods, if well used, can sometimes (fifth column) justify treating an individual with considerable concern, with subsequent assessment done “with prejudice” in the sense of being potentially extended
and including detention and aggressive questioning. That “sometimes”
should be understood as “occasionally,” however, and the methods typically have high false-alarm rates. The sixth columns uses “Yes, but . . .”
to indicate that yes, if a subject merits in-depth interrogation, most of
the methods can—as part of a more complex process with skilled security officers—be useful in obtaining confessions or information, but,
regrettably, they can also help generate false confessions. Abuse can
occur. The last column is crucial: None of the methods, except possibly
for analysis of textual or vocal content, are individually an adequate
xxxviii
Using Behavioral Indicators to Help Detect Potential Violent Acts
basis for arrest or conviction. Indeed, they may not be an adequate
basis for putting prejudicial information in a widely shared database
(e.g., “On such-and-such an occasion, the subject manifested facialexpression behaviors correlated with posing a security risk, although
other factors led to his being allowed to board the aircraft”).
This illustrates one of the many unresolved dilemmas. From a
purely detection perspective, and assuming a process for information
fusion, it would seem desirable to collect and share all kinds of fragmentary information of varied significance and credibility. However,
doing so could cause serious injustices to those affected and, in many
instances, would generate suspicions when none are scientifically warranted. It is instructive that, for almost a century, the FBI has maintained “raw files” on numerous subjects of observation, with important
instances of those files being misused (even though it can be argued
that this occurred rarely). How much more trouble would have been
created if analogous raw data had been widely shared? Such issues are
matters of degree, but no common agreement exists on what is and
is not reasonable. As a last example motivated by current discussions
in the news (as of January 2013), consider a teenager being treated
for symptoms of schizophrenia. What symptoms of violent tendencies
should trigger a report to authorities that would enter a sharable database, and with what balance of positive and negative consequences?
Such issues are profound. We made no attempt to resolve them except
that we see a major distinction between, on the one hand, using a
behavioral indicator as an increment of information in a detection
system seeking to identify, without further prejudice, which individuals merit more-than-usual scrutiny, and, on the other hand, using a
behavioral indicator to infer probable guilt or as the basis for arrest and
conviction. It is not accidental that the U.S. justice system has major
constraints on how methods such as polygraph techniques can be used.
Summary
xxxix
Conclusions
We found a number of important takeaways from our survey:
• Despite exaggerations found in commercial claims and the media,
there is current value and unrealized potential for using behavioral
indicators as part of a system to detect attacks. Unfortunately,
analytic quantification of that potential is poorly developed.
• “Operators” are often well ahead of the science base, which is
sometimes good and sometimes bad. It is very important that
programs build in and sustain objective evaluation efforts, despite
budgetary pressures and the tendency to see them as mere niceto-have items. The evaluations should be subjected to objective
peer review and adequate community scrutiny, although perhaps
within a classified domain. The Department of Defense and the
Intelligence Community have, for example, long used the federally funded research and development centers (FFRDCs),
national laboratories, National Academy of Sciences, and other
special panels for credible evaluations.
• Many serious problems and errors can be avoided by up-front
review of procedures by experts familiar with the subtleties of
detection and screening in conditions of high false-alarm rates
and low base rates. Although full validation of techniques may
take years (at a time when the dangers of attack are current),
existing knowledge can be used to avoide many problems that are
quite significant to privacy, civil liberties, travel and commerce.
• DHS and other security organizations are experimenting with
proposed methods—sometimes with laudable and ambitious scientific trials that have reported encouraging conclusions (which
are difficult to judge, however, without detailed access to data and
methods).
• Operators, their agencies, and the scientific community have not
done enough to understand how to mitigate the bad consequences
of detection systems, which invariably have false-alarm problems.
Much could be done.
xl
Using Behavioral Indicators to Help Detect Potential Violent Acts
• Information fusion is critical if behavioral indicators are to
achieve their potential. Fusion should occur not just within a
given method, but with heterogeneous information across activities and phases. Methods for accomplishing this are very poorly
developed. This said, it remains to be seen how much can realistically be accomplished. If the indicators being fused all have
very high false-alarm rates, the fused result may be more reliable
but still have a high false-alarm rate. Also, success in fusion will
depend on human skill in representing fuzzy, imperfect information.
• Information generation and retrieval, integration, and sensemaking will tax both automated methods (e.g., including for “big
data”) and perfecting human-machine interactions: Machines
can process vast amounts of data, but interpretation will continue
to depend on human expertise and judgment. An implications is
that “optimizing” should be for man-machine cooperation, not
automation.
• Very little research has been done to understand how much is
enough, but, subjectively, it seems that major improvements in
detection are plausible with networked real-time or near-real-time
integration of information. This would include further integrating (fusing) CIA and FBI information; proximate information at
checkpoints and fusion-center information; and criminal, commercial, security-related, and even whole-life information. What
can be accomplished is unclear, and developing a sharper understanding of payoff potential should be a priority task for objective
research and analysis.
• Such steps raise profound issues of privacy and civil liberties, but
the irony is that commercial organizations (and even political parties) are already far ahead in exploiting the relevant technologies
and forever changing notions of privacy.
• Investment decisions about individual technologies and methods
should be informed by structured portfolio-analysis approach
using something like the dimensions of Figure S.6.